Register an MCP server

Settings → AI Configuration → MCP Servers registers Model Context Protocol servers your agents can call. It needs the manage_integrations permission.

An MCP server is a service that exposes tools over a standard protocol. Registering one here is how an agent gets a capability that is specific to your company, without anybody writing an integration for it.

A realistic first server

The first MCP server most teams register is one they already run. Alpenglow’s is Summitline Analytics, a read-only service over their own warehouse, exposing three query tools: trip_metrics.query, trail_popularity.query, and retention.cohort.

That shape is the one to copy. It is internal, it is read-only, and its tools answer questions rather than change things. An agent that can read your metrics is useful on the first day; one that can write to them needs a conversation about autonomy first.

Register it

1. Click + Add Server.
2. Name and Description are what an agent sees, so say what it is for.
3. Server URL is where it lives.
4. Transport is HTTP, SSE, or stdio for a local process.
5. Status is Active or Inactive. Inactive keeps the registration and stops agents using it.

Scope: who may reach it

Scope is the access decision, and there are three:

Scope	Reaches
Tenant	Every agent in the workspace
Department	Agents in a role’s management chain
Position	The holder of one named role

Tenant is right for something like an analytics reader that any agent might want. Use Position when a server is genuinely one function’s tool, and give it that role’s id.

The tool whitelist

The whitelist names the tools an agent may call, comma separated. Leaving it empty allows everything the server exposes.

Fill it in. A server’s tool list is set by whoever runs the server, so an empty whitelist means “whatever that service decides to expose next month” rather than “the three tools I checked”. Naming three tools is a decision you can point at later.

Discovery lists, it does not grant

Discover asks the server what tools it has and shows them. It does not add them to the whitelist. That separation is the point: seeing a tool and allowing it are two different acts.

Test and watch

The detail pane holds two things worth returning to:

• Connection test answers whether the server is reachable right now, with the latency it took. Run it when an agent reports a tool failing.
• Usage statistics (24h) shows what agents actually called. A tool nobody has called in a month is a whitelist entry you can remove.

Next steps

• Agent skills: how a tool becomes something an agent knows when to use.
• The integration registry: the other way to give agents an outside capability.
• Roles: where a position scope points.